IBM A1000-125 - Assessment: AI Engineer Advanced Practice Exam: Hard Questions 2025

An AI engineer is developing a sentiment analysis model that consistently shows 92% accuracy on the training set but only 68% on the validation set. The model uses a deep neural network with 5 hidden layers. After analyzing the learning curves, the validation loss plateaus early while training loss continues to decrease. Which combination of techniques would MOST effectively address this issue?

A financial services company is deploying a Watson Natural Language Understanding service to extract entities and sentiment from customer feedback across multiple regions. They notice that the service performs well on US English data but poorly on feedback containing financial domain-specific terminology and multilingual code-switching between English and Spanish. What is the MOST comprehensive solution to improve model performance?

An AI engineer is implementing a reinforcement learning solution for an autonomous inventory management system. The agent must optimize order quantities while balancing carrying costs, stockout penalties, and supplier lead time variability. After 10,000 training episodes, the agent exhibits suboptimal behavior by consistently ordering maximum quantities regardless of demand patterns. The reward function is: R = revenue - (holding_cost × inventory) - (stockout_penalty × shortages). What is the MOST likely root cause and solution?

A healthcare organization is deploying multiple Watson Assistant instances across different departments. They need to implement a governance framework that ensures model versioning, A/B testing capabilities, audit trails for compliance, and automated rollback mechanisms. The solution must support deployment to both cloud and on-premises environments. Which architecture approach BEST meets these requirements?

An AI team is training a computer vision model to detect manufacturing defects from high-resolution images. They are experiencing memory overflow errors during training despite using a GPU with 16GB VRAM. The dataset contains 50,000 images at 4096×4096 resolution with 3 color channels. The current architecture uses a ResNet-50 backbone with batch size of 32. Which combination of techniques would allow training to proceed while MINIMIZING negative impact on model convergence?

A Watson Discovery instance is deployed for enterprise search across 500,000 legal documents. Users report that queries for specific case references (e.g., 'Brown v. Board 347 U.S. 483') return irrelevant results, while general concept searches work well. The current configuration uses default settings with no custom enrichments. Query logs show the tokenizer is splitting case references into separate terms. What is the MOST effective approach to resolve this issue?

An AI engineer is evaluating different loss functions for a multi-class image classification problem with 50 classes where the dataset exhibits severe class imbalance (largest class has 10,000 examples, smallest has 50). Initial models using standard cross-entropy loss achieve 94% accuracy but fail to detect minority classes. Which approach would MOST effectively improve minority class performance while maintaining overall model quality?

A deployed Watson Assistant chatbot for insurance claims is experiencing degraded performance with increased latency (3-5 seconds per response) and occasional timeout errors during peak hours. The assistant has 15 intents, 200 entities, 50 dialog nodes, and integrates with three external APIs for policy lookup, claims validation, and payment processing. Monitoring shows Watson Assistant API response times are normal (<200ms), but end-to-end user response times are poor. What is the MOST likely bottleneck and optimal solution?

An AI engineer is developing a Watson Visual Recognition custom classifier to identify 20 different types of industrial equipment from factory floor images. After training with 500 images per class, the model achieves 78% accuracy on the test set. Analysis reveals that visually similar equipment types (e.g., different models of forklifts) are frequently confused. The engineer has access to equipment metadata (manufacturer, model, year) but this isn't visible in images. What strategy would MOST effectively improve classifier performance?

A data science team is deploying a neural network model for real-time fraud detection that must process 10,000 transactions per second with latency under 50ms. The model has 3 million parameters and achieves 96% accuracy. Initial deployment using standard inference shows average latency of 120ms. The infrastructure uses cloud-based GPU instances. Which combination of optimization techniques would BEST meet the latency requirement while preserving model accuracy?